The prevailing story about AI in a disaster zone is that it belongs to governments and large NGOs: satellite imagery pipelines at FEMA, damage models at the World Bank, logistics dashboards at the UN. When twin earthquakes reportedly ripped through northern Venezuela last month, that story broke. The state was slow, the telecommunications backbone was in pieces, and the tools that actually reunited families, mapped collapsed buildings, and routed donations were built in three or four hours by programmers scattered between Buenos Aires, Santiago, Miami, and San Francisco. The technology was the same generation of large language models sitting on any developer’s laptop. The state was almost entirely absent.
That inversion — citizens with Claude and Replit outpacing a government with an emergency budget — is the story worth studying. It is not a triumphalist AI narrative. It is closer to a stress test of what happens when consumer-grade AI collides with a failing public sector, and what gets built, and what gets exposed, in the gap.
What actually happened in the first 24 hours
The earthquakes reportedly killed thousands of people, flattened homes and hospitals, and severed telecommunications infrastructure including fiber-optic cables and mobile network towers. Power failures compounded the connectivity problem. Social media, patchy and intermittent, filled with names.
A Venezuelan programmer living in Buenos Aires built Desaparecidos Terremoto Venezuela with a team of six. The site let people upload photos of missing relatives, cross-referenced them with facial recognition software, and did not require registration or an app download. In the first two days, it reportedly received more than 30,000 missing-person reports. The build would have taken roughly 24 hours without rest. Using Claude Opus, it took three.
In California, Venezuelan developer Samuel Mariña spun up Ayuda en Camino in four hours using Replit. It matched what people had — medicine, blankets, transport — against what people needed. A WhatsApp assistant handled queries for users who could not load the site because their connection was too weak.
An AI engineering company worked with a religious organisation to launch Somos Acompañamiento, which layered missing-person reports on top of hospital lists of survivors and the injured. The platform reportedly crossed 84,000 registrations. An AI system is also being built with the human rights group Laboratorio Migrante to watch for trafficking patterns in the displaced population.
From Santiago, Marianne Diaz Hernández — a digital security expert and the founder of the nonprofit Acceso Libre — built terremotove.com in about an hour, using Claude Code and Kobo Toolbox. By the time she published, at least three similar sites already existed. A women journalists’ network built its own verification tool, cross-checking hospital and shelter data. The Red de Sonadoras registry tracked people who had been located and treated.
Hundreds of AI-driven initiatives were reportedly activated worldwide in under a day. Three weeks later, these civilian-led platforms remain the primary source of crisis information in the country.
Why the tooling collapsed a two-day job into three hours
The interesting technical fact is not that AI wrote the code. It is which parts of the job it removed.
Setting up a disaster response website has always been technically straightforward and operationally miserable. The database schema for missing persons has to accommodate messy input: half-remembered addresses, misspelled surnames, photos taken from bad angles. The authentication flow has to be near-zero, because a grandmother on 2G cannot register an account. The frontend has to load on a phone with 200ms latency to a degraded network. Facial recognition has to run without shipping every uploaded image to a third-party API that may or may not be reachable. Each of these decisions used to require a specialist and a meeting.
Large language models compress the meetings. A developer describes the constraints in plain language — no login, WhatsApp fallback, low-bandwidth image compression, fuzzy matching on names — and gets a working scaffold in one pass. The developer still makes every architectural choice. The AI just executes the boilerplate that used to eat the hours between the choice and the deploy.
That is why the three-hour figure is credible rather than marketing. The developers were not writing novel algorithms. They were assembling well-understood patterns, faster, while making the humane decisions themselves.
The diaspora as a distributed emergency service
Venezuela has been exporting its skilled workforce for over a decade. Millions of Venezuelans now live abroad, and a significant slice of them are engineers. When the ground moved in Caracas and Aragua, the response infrastructure that mobilised was not physically inside the country. It was a diaspora graph on Twitter, Telegram, and WhatsApp, pinging Miami, Santiago, Buenos Aires, and the Bay Area within minutes.
This is a genuinely new pattern. Diaspora remittances have long been a fiscal lifeline for crisis states. What is different here is that the remittance is software, delivered in real time, by people who left precisely because the local institutions do not work. The developers who built these platforms have a personal grievance with the state they are substituting for. That gives the effort urgency, and it also gives it a permanent political edge that international NGOs try to avoid.
Diaz Hernández put it plainly: everything the state should be doing is being done by civil society, and people filled the state’s role however they could, with what they had.
The data-protection problem nobody wants to talk about mid-crisis
Here is where the story gets uncomfortable. The same platforms saving lives are also, in the language of privacy law, mass biometric collection points operating without a legal basis, in an authoritarian country, run by volunteers.
The Desaparecidos Terremoto site alone ingested 30,000 photos of faces in 48 hours, cross-referenced by facial recognition. Somos Acompañamiento holds identifying information on 84,000 people. Ayuda en Camino, being a matching platform, knows who has what and who needs what, which in a politicised aid environment is itself sensitive.
Digital rights advocates have noted that AI can complement family reunification and aid distribution and make them more efficient, but that it should never replace the state’s legal responsibility or its accountability mechanisms. Data collected in Venezuela needs safeguards stronger than usual, and treating privacy as a post-crisis concern is a mistake.
That warning is not abstract. Desaparecidos Terremoto has reportedly already come under cyberattack, prompting the team to harden the platform’s security. Whoever launched those attacks now has a target list of people looking for missing relatives — a database whose future custody is entirely unclear. The question of what happens to the underlying data is unresolved.
Facial recognition datasets are especially thorny. A photo uploaded during an earthquake to find a grandmother becomes, on a different server three years later, a training set. Or a policing tool. Or leverage. The consent given at upload — implied, hurried, sometimes given by a neighbour rather than the person pictured — does not scale forward in time.
The state accountability question
There is a subtler critique buried in the praise. When civil society plugs the gap this effectively, it can inadvertently underwrite the failure it is trying to compensate for.
If a functional missing-persons registry appears in three hours and the government has to do nothing to produce it, the political cost of state failure drops. The families still suffer. The building still collapsed. But the immediate visible failure — no one is coordinating anything — gets absorbed by volunteers, and the acting government, whose response to the criticism has been to reject it, keeps its footing.
Cybersecurity experts have noted the operational costs: information across the citizen platforms is disaggregated, sometimes incorrect, sometimes duplicated, and privacy risks compound on top. The state should be centralising efforts and protecting people’s privacy — functions the civilian sites cannot legitimately perform even when they perform them well.
None of this is an argument against what these developers did. It is an argument for keeping the accountability ledger honest. The UN’s own updates on the Venezuelan situation lean heavily on data streams that ultimately originate in citizen platforms. When international aid uses volunteer registries as source of truth, it lends them a legitimacy without providing the governance they lack.
What generalises beyond Venezuela
The uncomfortable finding is that the Venezuela pattern is likely to repeat, and quickly.
Three ingredients made this response possible: a large technically skilled diaspora, cheap and capable general-purpose AI, and a state either unwilling or unable to run a modern emergency response. That combination now exists in dozens of countries. Lebanon has it. Sudan has it. Myanmar has it. Parts of Central America and the Sahel have it. Any future flood, earthquake, or blackout in those places will trigger the same reflex — diaspora developers spinning up platforms in hours, running on consumer AI subscriptions, holding sensitive data with no governance layer.
Some of the Venezuelan developers have already generalised. One developer who built a missing-persons registry called Civis is repurposing it as an early-warning system for floods and landslides, using open-source satellite data and rainfall forecasts to send SMS and WhatsApp alerts to rural communities that currently receive nothing from any official channel. That is a citizen-built meteorological warning service. It will save lives. It also lives entirely outside any regulatory framework for public alerts.
A tool, not a substitute
The most honest line in the entire account: AI is a tool, it is not a substitute for people, and the people are the ones who make these tools valuable.
That framing is useful because it cuts against two lazy readings of the story. The first reading is that AI has now made state disaster response obsolete — that any competent developer with Claude can outperform a national emergency agency, so why bother funding the agency. The second reading is that AI is dangerous in humanitarian contexts because it collects biometrics and moves fast, so it should be constrained. Both readings miss the actual variable.
The variable is trust. The Venezuelan citizens uploading photos of their missing relatives to Desaparecidos Terremoto trusted a programmer in Buenos Aires more than they trusted their own government. That trust is the load-bearing element. The AI just made it possible to convert that trust into working software in an afternoon rather than a week. Take away the trust and the same tools produce nothing useful. Take away the tools and the trust still exists, but the response is slower and smaller.
Which is why the durable lesson has less to do with AI and more to do with what a country’s civic infrastructure looks like on an ordinary Tuesday. When Finland, as covered in a recent piece on its education system, spent decades building deep institutional trust, it did so specifically to make the state usable in a crisis. Venezuela’s citizens have built the opposite pattern — trust in each other and in the diaspora, in explicit substitution for the state. Both are functioning systems. Only one is stable.
What to watch over the next three months
The Desaparecidos Terremoto team has committed to keeping the site online for roughly three months, and to keeping it available for future disasters. Somos Acompañamiento is expanding into trafficking monitoring. Ayuda en Camino is still routing donations. Terremotove.com continues to log damage reports.
Three questions will decide whether this becomes a template or a cautionary tale. First, what happens to the data. Whether it gets deleted, encrypted, handed to a legitimate custodian, or quietly leaked will shape whether future disaster victims trust the next citizen platform. Second, whether the international humanitarian system treats these platforms as infrastructure to integrate or as a stopgap to route around. Right now the formal aid response is arriving weeks late while the citizen registries remain the main source of truth — and the people who built them are clear that this is not how it should stay. The state, in the words of one cybersecurity consultant, should be the one centralising efforts and protecting people’s privacy; nobody involved is arguing that a volunteer in Santiago is a permanent substitute for that. Third, whether the reflex institutionalises. The same developers are already pointing their tools at the next crisis — flood alerts, landslide warnings, trafficking surveillance — domains that normally sit behind a public agency and a regulatory framework, and that will now be served, at least in some places, by a small team and an API key.
None of those questions has a clean answer yet. What is already settled is the demonstration itself: a modern emergency response was stood up in an afternoon by people the state had driven out, and it held for three weeks while the state did not. That is either the most encouraging thing to happen to disaster relief in a decade or the most quietly destabilising, and which one it turns out to be will be decided not by the models but by what the people holding the data choose to do next.
